A manifesto for recursive discovery

Biology must learn from the world.

Not another model. A system that earns intelligence through experiment.

Begin reading ↓

Perseverance’s coded parachute · NASA/JPL-Caltech [66]

Biology does not lack ideas. It lacks a reliable return path from idea to evidence.

We can predict structures, generate sequences, rank molecules and draft protocols at extraordinary speed. Yet the decisive fact still lives outside the model: in a cell, an assay, a tissue, a patient. Intelligence that never meets that world can only rearrange what is already known.

Models predict. Chemists build. Laboratories test. Instruments read. Scientists interpret. These parts live in different places, formats and timescales. The handoffs erase context; negative results disappear; every new program begins too close to zero.

Self-driving laboratories show that a model can select experiments, a laboratory can execute them and measurements can update the next choice.5, 6, 7, 8, 9, 10, 11 The important breakthrough is not automation by itself. It is the closure of the loop.

The thesis

Prediction becomes discovery only when physical evidence returns.

Read → write → learn → compoundThe operating loop, rendered as one continuous system.

Build a physical reinforcement loop for biology.

A scientific agent proposes an intervention. A qualified laboratory executes a standardized experiment in living systems. Functional and multimodal measurements read the response. An uncertainty-aware selector chooses what should happen next. The complete trajectory returns to memory.

01

Read

Measure biological state, function, safety and response.

02

Write

Place a precise intervention into a living system.

03

Learn

Choose the experiment that changes the decision.

04

Compound

Return every result—positive, negative or failed—to the system.

The unit of progress is not a model, an assay or a molecule. It is a decision that becomes better because evidence returned to the loop.

Begin narrow enough to be wrong.

Dellvi’s first node is kidney-directed siRNA delivery: one tissue, one target cell, one payload class, one locked functional assay and two prospective experimental rounds.

Delivery determines where a programmable payload goes, which cell internalizes it, whether it escapes the endosome, how long it acts and which other cells receive it. It is not packaging around the drug; it is a large part of the therapeutic index.24, 25, 26, 27, 28

The kidney proximal tubule is clinically important, specialized for high endocytic flux and difficult enough to expose weak reasoning. Megalin, cubilin and Dab2 participate in proximal-tubule uptake, while recent work demonstrates kidney targeting with chemically modified carbohydrate conjugates.29, 30, 31, 32

12Mechanistically diverse candidates in round one
6Prospective choices frozen before round two
2Payloads to test whether the learning transfers
1Traceable decision history from question to result

Organoids and flow-conditioned tubules are escalation layers, not decorative complexity. They enter only when they resolve an uncertainty that a cheaper model cannot.33, 34, 35, 36, 37, 38

An operating belief

Dare mighty things.

NASA/JPL-Caltech · Mars Decoder Ring, PIA24431 [66]

When Perseverance descended into Jezero Crater, its parachute carried more than a rover. The orange-and-white gores encoded a message.

The inner rings spell DARE MIGHTY THINGS. The outer ring encodes the coordinates of NASA’s Jet Propulsion Laboratory. Systems engineer Ian Clark designed the pattern.66 The object that slowed a spacecraft in the Martian atmosphere also carried the culture that built it.

The metaphor matters because the artifact was not merely brave. It was tested. The 72-foot canopy had to inflate in a Mach 2 wind in roughly half a second. NASA tested the system repeatedly under Mars-relevant conditions before entrusting it with the mission.67

Ambition is not the opposite of rigor. The only ambition worth trusting is the kind that survives contact with reality.

Borrow models. Rent capacity. Own the feedback loop.

Dellvi should be a neutral delivery foundry. Partners bring therapeutic objectives and payloads. Dellvi designs and validates delivery systems. Partners take nominated candidates through IND-enabling work, clinical development and commercialization.

The partner owns its payload, target and therapeutic program. Dellvi owns carrier platform IP, the experimental operating system and the generalized learning created across programs. The model can be replaced. The decision history cannot.

Open foundation models should supply broad priors, representations and design proposals.12, 13, 14, 15, 16, 17, 18, 19, 20 Dellvi’s proprietary layer is the calibrated surrogate trained on its own physical outcomes—and the system that knows which experiment is worth buying next.

Partner brings

Objective, payload, product constraints, downstream development.

Dellvi compounds

Carrier IP, assay truth, experimental memory, selection policy.

We will not call fluorescence delivery.

We will not call a retrospective fit learning.

We will not call outsourced experiments a platform.

We will earn every claim prospectively.

The decision history is the asset.

A flat spreadsheet of final values is not memory. Every result must connect the objective, admissible design space, material, biological context, protocol, raw measurement, quality state, cost, interpretation and next decision.

Negative results remain useful. They shrink the design space, expose broken assays, improve uncertainty calibration and prevent the next program from starting at zero.

The practical score is validated learning per dollar and validated learning per week—not model size, automation theatre or the number of experiments performed.

Turn cost advantage into learning advantage.

India can connect exceptional chemistry, biology, software, clinical and manufacturing capability into a shared learning network. The opportunity is not cheaper labour. It is more high-quality, decision-changing experiments per unit of capital.

Biocon and Syngene show that globally relevant biotechnology and research infrastructure can be built from India; national biomanufacturing initiatives add further momentum.46, 47, 48, 49 Dellvi’s job is to turn that distributed capability into a versioned, observable system where raw data and provenance always return.

Generalize only after the narrow loop works.

  1. R0

    Experimental contract

    One experiment reconstructable end to end.

  2. R1

    Assay verifier

    A qualified human proximal-tubule assay with stable controls.

  3. R2

    Closed loop

    A prospective round-two selection compared with expert and diversity baselines.

  4. R3

    Platform transfer

    The same delivery handle works with a second payload and higher-fidelity model.

  5. R4

    Partner product

    A paid Kidney Delivery Challenge with a decision-grade evidence room.

  6. R5

    Translation

    In-vivo biodistribution, functional pharmacodynamics and an interpretable safety margin.

Platform learning can reduce repeated work, but it cannot erase product-specific manufacturing, safety and regulatory evidence.50, 51, 52, 53

The long view

Build one loop that learns, transfers and earns revenue. Then repeat.

Dellvi exists to give intelligent systems a way to ask the physical world new questions—and to earn their generality through experimentally verified action.

Read the evidence ↓

Source map

References

Scientific, technical and commercial references are drawn from the Recursive Discovery Factory build blueprint supplied for this manifesto. Company-reported results remain company claims. NASA references were checked against JPL source pages.

  1. 01Lila Sciences. Tech: Building Scientific Superintelligence.
  2. 02Lila Sciences. From Prompt to CAR-T: Inside Lila’s RNA Discovery Engine.
  3. 03Kompa B. Building the Agent-Driven Era of Science with NVIDIA BioNeMo Agent Toolkit.
  4. 04Lila Sciences. Solutions.
  5. 05Canty RB, Abolhasani M. The past, present and future of self-driving laboratories. Nature Reviews Chemistry. 2026.
  6. 06Roch LM et al. ChemOS: Orchestrating autonomous experimentation. Science Robotics. 2018.
  7. 07Burger B et al. A mobile robotic chemist. Nature. 2020.
  8. 08MacLeod BP et al. Self-driving laboratory for accelerated discovery of thin-film materials. Science Advances. 2020.
  9. 09Rapp JT, Bremer BJ, Romero PA. Self-driving laboratories to autonomously navigate the protein fitness landscape. Nature Chemical Engineering. 2024.
  10. 10Borkowski O et al. Large-scale active-learning-guided exploration for in vitro protein production optimization. Nature Communications. 2020.
  11. 11Wu Z et al. Active learning-assisted directed evolution. Nature Communications. 2025.
  12. 12Watson JL et al. De novo design of protein structure and function with RFdiffusion. Nature. 2023.
  13. 13Dauparas J et al. Robust deep learning-based protein sequence design using ProteinMPNN. Science. 2022.
  14. 14Abramson J et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature. 2024.
  15. 15Brixi G et al. Genome modelling and design across all domains of life with Evo 2. Nature. 2026.
  16. 16Hayes T et al. Simulating 500 million years of evolution with a language model. Science. 2025.
  17. 17Chen J et al. Interpretable RNA foundation model from unannotated data for RNA structure and function predictions. 2022.
  18. 18Penić RJ et al. RiNALMo: general-purpose RNA language models can generalize well on structure prediction tasks. Nature Communications. 2025.
  19. 19Cui H et al. scGPT: toward building a foundation model for single-cell multi-omics using generative AI. Nature Methods. 2024.
  20. 20Theodoris CV et al. Transfer learning enables predictions in network biology. Nature. 2023.
  21. 21Sample PJ et al. Human 5′ UTR design and variant effect prediction from a massively parallel translation assay. Nature Biotechnology. 2019.
  22. 22Zhang H et al. Algorithm for optimized mRNA design improves stability and immunogenicity. Nature. 2023.
  23. 23Rurik JG et al. CAR T cells produced in vivo to treat cardiac injury. Science. 2022.
  24. 24Setten RL, Rossi JJ, Han SP. The current state and future directions of RNAi-based therapeutics. Nature Reviews Drug Discovery. 2019.
  25. 25Nair JK et al. Multivalent GalNAc-conjugated siRNA localizes in hepatocytes and elicits robust gene silencing. JACS. 2014.
  26. 26Springer AD, Dowdy SF. GalNAc-siRNA conjugates: leading the way for delivery of RNAi therapeutics. Nucleic Acid Therapeutics. 2018.
  27. 27Gilleron J et al. Image-based analysis of LNP-mediated siRNA delivery, trafficking and endosomal escape. Nature Biotechnology. 2013.
  28. 28Wittrup A et al. Visualizing lipid-formulated siRNA release from endosomes and target gene knockdown. Nature Biotechnology. 2015.
  29. 29Huang Y et al. Physiological principles underlying the kidney targeting of renal nanomedicines. Nature Reviews Nephrology. 2024.
  30. 30Prins J et al. RNA-based therapeutic opportunities for the treatment of kidney diseases. Nature Reviews Nephrology. 2026.
  31. 31Rbaibi Y et al. Megalin, cubilin and Dab2 drive endocytic flux in kidney proximal tubule cells. Molecular Biology of the Cell. 2023.
  32. 32Kumar V et al. Efficient and selective kidney targeting by chemically modified carbohydrate conjugates. Molecular Therapy. 2024.
  33. 33Takasato M et al. Kidney organoids from human iPS cells contain multiple lineages and model human nephrogenesis. Nature. 2015.
  34. 34Morizane R et al. Nephron organoids derived from human pluripotent stem cells model kidney development and injury. Nature Biotechnology. 2015.
  35. 35Homan KA et al. Flow-enhanced vascularization and maturation of kidney organoids in vitro. Nature Methods. 2019.
  36. 36Freedman BS et al. Physiology assays in human kidney organoids. American Journal of Physiology Renal Physiology. 2022.
  37. 37Ma C et al. Efficient proximal tubule-on-chip model from hiPSC-derived kidney organoids for functional analysis of renal transporters. 2024.
  38. 38Jang KJ et al. Human kidney proximal tubule-on-a-chip for drug transport and nephrotoxicity assessment. Integrative Biology. 2013.
  39. 39Bryant DH et al. Deep diversification of an AAV capsid protein by machine learning. Nature Biotechnology. 2021.
  40. 40Ogden PJ et al. Comprehensive AAV capsid fitness landscape reveals a viral gene and enables machine-guided design. Science. 2019.
  41. 41Schuhmacher MK et al. Creating bottom-up RNA transfer vehicles from synthetic protein assemblies. Nature. 2026.
  42. 42Judo Bio. Launch and $100 million financing for kidney-directed oligonucleotide therapies. 2024.
  43. 43Judo Bio. The STRIKE platform enables kidney-selective gene silencing via megalin-mediated uptake. OTS poster. 2025.
  44. 44Dyno Therapeutics. Partnerships.
  45. 45Dyno Therapeutics. New strategic partnership with Roche. 2024.
  46. 46Biocon. Our legacy.
  47. 47Syngene International. Annual Report 2025: Strategic priorities.
  48. 48BIRAC. BioE3 policy and biomanufacturing initiative.
  49. 49BIRAC. India Bioeconomy Report 2026.
  50. 50US FDA. CMC information for human gene therapy INDs. 2020.
  51. 51US FDA. Platform Technology Designation Program for Drug Development. Draft guidance. 2024.
  52. 52US FDA. Drug Master Files: Guidelines.
  53. 53US FDA. FDA requests Sarepta suspend distribution of Elevidys and revokes AAVrh74 platform technology designation. 2025.
  54. 54Arc Institute. Evo 2 official code, model access and training resources.
  55. 55RosettaCommons. RFdiffusion official repository.
  56. 56Dauparas J et al. ProteinMPNN official repository.
  57. 57Wohlwend J et al. Boltz official biomolecular-interaction models repository.
  58. 58Wang B laboratory. scGPT official repository.
  59. 59Theodoris CV. Geneformer official model repository.
  60. 60LBCB. RiNALMo official implementation and weights.
  61. 61A*STAR / ML4Bio. RNA-FM official repository.
  62. 62RDKit contributors. RDKit official repository.
  63. 63Chemprop contributors. Chemprop official repository.
  64. 64PyTorch. BoTorch official repository.
  65. 65Cornellius GP. GPyTorch official repository.
  66. 66NASA/JPL-Caltech. Mars Decoder Ring: the binary message in Perseverance’s parachute. 2021.
  67. 67NASA/JPL-Caltech. Testing Proves Its Worth With Successful Mars Parachute Deployment. 2021.